TensorFlow1（二）文件读取_tensorflow根据文件夹读取数据

我们首先说一下文件读取的流程（分别讨论文本文件、图片文件以及二进制文件）：

 1、构造文件名队列

file_queue= tf.train.string_input_producer(string_tensor,shuffle = True)

2、读取与解码

文本：

        读取：tf.TextLineReader()

        解码：tf.decode_csv()

图片：

        读取：tf.WholeFileReader()

        解码：tf.image.decode_jepg(contents)

                   tf.image.decode_png(contents)

二进制文件：

         读取：tf.FixedLengthRecordReader(record_bytes)
         解码：tf.decode_raw()

3、批处理队列

 tf.train.batch(tensors, batch_size, num_threads = 1, capacity = 32, name=None)

手动开启线程：coord = tf.train.Coordinator()

开启会话：tf.train.start_queue_runners(sess=None, coord=None)

回收线程：coord.request_stop()   coord.join(threads)

实例一：狗的图片读取

我这里使用的是100张狗的图片，图片类型是jpg，因为每个图片的规格都是不同的，因此加入了reshape的步骤。

import os
 
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2'
import tensorflow as tf
 
# 图片数据
# 图片 - 数值（三维数组shape（图片长度、图片宽度、图片通道数））
 
# 图片三要素
# 灰度图 [长,宽,1] 每一个像素点[0,255]
# 彩色图 [长,宽,3] 每一个像素点[0,255]
 
def picture_read(file_list):
    # 1.构造文件名队列
    file_queue = tf.train.string_input_producer(file_list)
    # 2.读取与解码
    reader = tf.WholeFileReader()
    # key文件名 value一张图片的原始编码形式
    key,value = reader.read(file_queue)
    print(key,value)
    # 解码阶段
    image = tf.image.decode_jpeg(value)
    print(image)
    # 图像的形状、类型修改
    image_resized = tf.image.resize_images(image,size = [200,200])
    print(image_resized)
    #静态形状修改
    image_resized.set_shape(shape=[200,200,3])
    # 3.批处理
    image_batch = tf.train.batch([image_resized],batch_size=100,num_threads=1,capacity=100)
    print("image_batch:",image_batch)
    # 开启会话
    with tf.Session() as sess:
        # 开启线程
        # 线程协调员
        coord = tf.train.Coordinator()
        threads = tf.train.start_queue_runners(sess=sess,coord=coord)
        key_new,value_new,image_new,image_resized,image_batch = sess.run([key,value,image,image_resized,image_batch])
        print("key_new:",key_new)
        # print("value_new:",value_new)
        print("image_new",image_new)
        print("image_resized",image_resized)
        print("image_batch",image_batch)
 
        # 回收线程
        coord.request_stop()
        coord.join(threads)
 
if __name__ == "__main__":
    filename = os.listdir("./dog")
    # print(filename)
    # 拼接路径 + 文件名
    file_list = [os.path.join("./dog/",file) for file in filename]
    # print(file_list)
    picture_read(file_list)

示例二：读取二进制文件

本示例采用的是CIFAR-10 binary version (suitable for C programs)

import os
 
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2'
import tensorflow as tf
 
class Cifer(object):
    def __init__(self):
        # 初始化操作
        self.height = 32
        self.width =  32
        self.channels = 3
 
        # 字节数
        self.image_bytes = self.height * self.width * self.channels
        self.label_bytes = 1
        self.all_bytes = self.label_bytes + self.image_bytes
 
    def read_and_decode(self,file_list):
        # 构造文件名队列
        file_queue = tf.train.string_input_producer(file_list)
 
        # 读取与解码
        # 读取阶段
        reader = tf.FixedLengthRecordReader(self.all_bytes)
        # key文件名 value一个样本
        key,value = reader.read(file_queue)
        print("key:",key)
        print("value:",value)
        # 解码阶段
        decoded = tf.decode_raw(value,tf.uint8)
        print("decoded:",decoded)
        # 将目标值和特征值切片切开
        label = tf.slice(decoded,[0],[self.label_bytes])
        image = tf.slice(decoded,[self.label_bytes],[self.image_bytes])
        print("label:",label)
        print("image:",image)
        # 调整图片形状
        image_reshaped = tf.reshape(image,shape=[self.channels,self.height,self.width])
        print("image_reshaped:",image_reshaped)
        # 转置，将图片的顺序旋转为height,width,channels
        image_transposed = tf.transpose(image_reshaped,[1,2,0])
        print("image_transposed",image_transposed)
        # 调整图像类型
        image_cast = tf.cast(image_transposed,tf.float32)
        # 批处理
        label_batch,image_batch= tf.train.batch([label,image_cast],batch_size=100,num_threads=1,capacity=100)
        print("label_batch",label_batch)
        print("image_batch",image_batch)
        # 开启会话
        with tf.Session() as sess:
            # 开启线程
            coord = tf.train.Coordinator()
            threads = tf.train.start_queue_runners(sess=sess,coord=coord)
 
            key_new,value_new,decoded_new,label_new,image_new,image_reshaped_new,image_transposed_new = sess.run([key,value,decoded,label,image,image_reshaped,image_transposed])
            label_value,image_value = sess.run([label_batch,image_batch])
            print("key_new:", key_new)
            # print("value_new:", value_new)
            print("decoded_new",decoded_new)
            print("label_new",label_new)
            print("image_new",image_new)
            print("image_reshaped_new",image_reshaped_new)
            print("image_transposed_new",image_transposed_new)
            print("label_batch",label_batch)
            print("image_batch",image_batch)
            # 回收线程
            coord.request_stop()
            coord.join(threads)
        return image_value,label_value
if __name__ == "__main__":
    file_name = os.listdir("./cifar-10-binary/cifar-10-batches-bin")
    # print("file_name:",file_name)
    # 构造文件名路径列表
    file_list = [os.path.join("./cifar-10-binary/cifar-10-batches-bin/",file) for file in file_name if file[-3:] == 'bin']
    # print("file_list:",file_list)
    # 实例化Cifar
    cifar = Cifer()
    image_batch,label_batch = cifar.read_and_decode(file_list)